The Artificial Intelligence Dictionary
Want to learn everything about artificial intelligence?
Whether you’re a beginner, an expert, or just curious, our AI dictionary is regularly updated to help you understand this field. Click on the word to read the complete definition.
Agentic AI : Agentic AI is an artificial intelligence system capable of acting autonomously, making decisions, and executing complex tasks in a given environment, often with a specific objective, dynamically adapting to changes without constant human intervention.
AI Slop: Very low-quality content, massively generated by AI models (especially the cheapest or poorly prompted versions), often full of clichés, factual errors, visual artifacts, hollow phrasing, and a complete lack of soul or real intent.
Alan Turing: A British mathematician and cryptanalyst, Alan Turing is considered the father of modern computer science. His work paved the way for artificial intelligence.
Algorithm: An algorithm is a step-by-step series of instructions designed to solve a problem or accomplish a task. It is essential in computer science for data processing and function execution.
Anthropic: A company founded in 2021 by former OpenAI researchers, Anthropic develops artificial intelligence models, such as Claude, with a focus on safety and ethics through a ‘Constitutional AI’ approach. Despite promising innovations and massive financial support (Amazon, Google), it faces controversies, particularly regarding copyright and military applications.
Apertus: A large open-source language model developed in Switzerland by EPFL, ETH Zurich, and CSCS. Designed for the common good, it stands out for its multilingualism, transparency, and regulatory compliance, offering an ethical and secure alternative for local applications.
API (Application Programming Interface): An API allows two computer systems to communicate with each other. In the context of artificial intelligence, APIs provide access to powerful models like those of ChatGPT, for example.
Self-supervised learning: A machine learning method where a model learns representations from raw data without explicit annotations, by generating its own prediction tasks, such as guessing masked words in a text or associating modified versions of an image.
Supervised learning: Supervised learning is a machine learning method where a model is trained on a labeled dataset, meaning data for which the desired outcomes are known. The goal is to learn to predict outcomes from new data by relying on the relationships identified during training.
AutoML (Automated Machine Learning): A process that simplifies the creation of artificial intelligence models by automating steps such as algorithm selection, data preprocessing, and hyperparameter optimization, thus enabling users, even without technical expertise, to develop high-performing models.
Backbone: The main part of an artificial intelligence model, usually a large pre-trained neural network (like ResNet, ViT, BERT, Llama, etc.), which extracts basic features from raw data (images, text, audio…) and serves as a reusable foundation for specialized downstream tasks via fine-tuning or adaptation.
Algorithmic Bias in AI: Algorithmic bias in artificial intelligence refers to an error where an AI system produces unfair or discriminatory results, often by reproducing prejudices present in the data it was trained on.
Big Data: Big Data refers to the large, varied, and constantly evolving datasets that cannot be processed efficiently by traditional methods. It involves using advanced technologies to collect, store, analyze, and extract meaningful information from this data to make informed decisions.
Quantum Computing (in AI): Quantum computing in artificial intelligence could lead to quantum AI. This technology uses the principles of quantum mechanics to perform complex calculations faster and more efficiently than traditional computers, thus opening new perspectives for the development of advanced algorithms.
Chatbot: A chatbot is a computer program that simulates human conversation, using AI to understand and respond to interactions. ChatGPT is an advanced example of a chatbot, capable of generating complex text and conversing fluently.
Computer Vision: Computer vision is a field of AI that enables computers to extract meaningful information from images and videos. The goal is to make machines capable of understanding and acting upon the visual world.
DALL-E: An AI model developed by OpenAI that generates images from text descriptions, combining concepts from the name “DALL·E” (a play on Salvador Dalí and Pixar’s WALL-E character) and advanced machine learning techniques.
Data Augmentation: Data augmentation is a technique that involves enriching a dataset by generating new data from existing data through transformations (rotations, substitutions, noise addition, etc.), in order to improve the performance and robustness of AI models.
Dataset: An organized collection of data, such as texts, images, videos, or numbers, used to train, validate, or test artificial intelligence models in tasks like image recognition or language processing. Its quality, diversity, and precise labeling are essential for algorithms to learn effectively and generalize their predictions to new situations.
Deep Learning: Deep Learning is a subfield of Machine Learning that uses artificial neural networks with multiple layers to model complex data representations. It is particularly effective for processing tasks such as image recognition, natural language processing, and other applications requiring a deep understanding of data.
Edge AI: Edge AI refers to artificial intelligence deployed directly on local devices, such as smartphones, cars, or sensors, to process data in real time without relying on the cloud. This improves speed, privacy, and energy efficiency.
Embedding : An embedding is a numerical representation of data (such as a word, an image, or a user) in the form of a vector, allowing artificial intelligence to analyze and compare it efficiently.
AI Ethics: AI ethics encompasses questions and principles aimed at ensuring that AI systems are developed and used responsibly, transparently, and respectfully of human rights. This field addresses issues such as bias, privacy, security, and the societal impact of AI.
Explainable AI: “Explainable AI” refers to a set of techniques and approaches in artificial intelligence aimed at making the decisions and reasoning processes of AI models understandable and transparent to humans, by clearly explaining the steps and criteria that lead to a conclusion or prediction.
Feature Engineering: Also known as “Ingénierie des caractéristiques” (Feature Engineering), this is the process of creating, transforming, and selecting features (variables) from raw data to improve the performance of artificial intelligence and machine learning models.
Fine-tuning: Fine-tuning is an AI technique that involves adjusting a pre-trained model on specific data to improve its performance on a targeted task. This optimizes efficiency while leveraging the model’s general knowledge.
Foundation Model: A foundation model is a large artificial intelligence model, pre-trained on enormous volumes of data, designed to serve as a base for many different tasks such as language processing, image analysis, or content generation.
GAIA Benchmark: GAIA is a benchmark (reference test) that evaluates general artificial intelligence through 466 practical questions, simple for humans but difficult for AIs.
GANs: Generative Adversarial Networks, or GANs, are a deep learning architecture where two neural networks compete: the generator creates synthetic data, while the discriminator evaluates its authenticity. This competition improves the quality of the generated data, and GANs are used for creating images, videos, and other synthetic content.
Gemini: A family of multimodal artificial intelligence models developed by Google, capable of processing and generating text, images, videos, and other data. Launched in 2023, it aims to compete with cutting-edge AIs while integrating into the Google ecosystem for various applications.
Geoffrey Hinton: AI pioneer and father of deep learning. A professor in Toronto, he received the 2018 Turing Award and has been warning about the ethical risks of AI since 2023.
GPT: An acronym that appeared in 2018, “Generative Pre-trained Transformer” is a natural language processing model developed by OpenAI, designed to generate coherent and contextually relevant text.
GPU : A GPU, or Graphics Processing Unit, is a powerful processor that excels at simultaneous calculations, widely used in artificial intelligence to boost the training and use of models, such as those in Deep Learning or computer vision.
Large Language Model (LLM): This is a generative artificial intelligence system, trained on vast textual data, capable of understanding and producing natural language for tasks such as conversation, translation, or text generation.
Grok: Launched online in 2023, Grok is an artificial intelligence developed by Elon Musk with his company X-AI. Grok is available on the X platform (formerly Twitter).
AI Hallucination: An AI hallucination is an error where an artificial intelligence model generates false or invented information, often convincingly, due to a misinterpretation of data or a lack of context.
AI ACT: The AI Act is the European regulation on artificial intelligence. It is a legislative proposal by the European Union aimed at regulating the use of artificial intelligence. Its main objective is to ensure the safety, transparency, and ethics of AI systems while fostering innovation.
Ilya Sutskever: Co-founder of OpenAI and creator of Safe Superintelligence Inc., Ilya Sutskever is an AI pioneer known for his breakthroughs with AlexNet and his commitment to developing safe superintelligence. His career combines technical innovation and ethical reflection to shape the future of artificial intelligence.
Imitation Learning: An AI technique where a model learns to reproduce behavior by observing examples performed by a human or an expert, rather than learning through trial and error.
Inference : In AI, this is the process by which a trained model uses its knowledge to produce predictions or outputs on new data. It is the application phase of AI.
Artificial Intelligence: Artificial intelligence refers to the simulation of human intelligence processes by computer systems, including learning, reasoning, and self-correction.
General Artificial Intelligence (AGI): General Artificial Intelligence (or Super AI) is a type of AI capable of understanding, learning, and applying knowledge across a variety of tasks, similar to human intelligence.
Generative Artificial Intelligence: Generative AI is a technology that creates original content, such as text, images, or music, by drawing inspiration from the data it has been trained on.
Artificial Superintelligence (ASI): ASI, or Superintelligence, refers to AI that far surpasses human intellectual capabilities in all areas: creativity, problem-solving, decision-making, etc. Unlike AGI (Artificial General Intelligence), which equals human intelligence, ASI would be exponentially more powerful, capable of self-improvement and solving complex problems on an unimaginable scale, raising major ethical and existential issues.
Judea Pearl : Israeli-American computer scientist born in 1936, father of Bayesian networks, creator of the modern framework for causal inference, recipient of the 2011 Turing Award, a leading author in artificial intelligence, and a thinker whose work has transformed statistics, medicine, economics, and the philosophy of science.
Llama: A series of open-source AI models developed by Meta AI since 2023. Initiated by FAIR, it evolves from Llama 1 to Llama 4, supporting research and innovation.
Low-Rank Adaptation (LoRA): A fine-tuning method that freezes the pre-trained model and only trains small, low-rank matrices added in parallel, drastically reducing parameters (0.1-1%) while maintaining performance close to full fine-tuning and minimal memory footprint.
Machine Learning: Machine Learning is a sub-discipline of artificial intelligence that allows computer systems to learn from data and improve their performance without being explicitly programmed.
Marvin Minsky: An American pioneer of artificial intelligence, co-founder of the MIT AI Lab, and recipient of the 1969 Turing Award. Inventor of the confocal microscope, theorist of neural networks, and author of The Society of Mind, he revolutionized our view of the human mind by comparing it to a “society” of simple agents. An eccentric and multidisciplinary visionary, he remains one of the founding fathers of modern AI.
Attention Mechanism: A technique in artificial intelligence allowing a model to prioritize certain parts of input data according to their relevance for a task, optimizing contextual understanding and processing efficiency.
Mistral AI: A French startup founded in 2023, which develops artificial intelligence models, particularly large language models (LLMs), with the goal of making them open and ethical.
Diffusion Models: Diffusion models are generative AI algorithms that create data (like images) by modeling a progressive transformation process: they start from random noise and refine it step by step to generate a coherent output, relying on probabilistic principles.
NLP (Natural Language Processing): Natural Language Processing is a branch of artificial intelligence that enables machines to understand, interpret, and generate human language naturally.
John McCarthy: Considered “the father of artificial intelligence,” John McCarthy was an American computer scientist, a pioneer in the field of AI, known for organizing the first AI conference at Dartmouth College in 1956. He also developed the LISP programming language, which became a fundamental tool for AI research.
OpenAI : Founded in 2015, OpenAI is an American artificial intelligence research and development company. It is known for its major advancements in language and image generation models, notably with innovations like GPT (ChatGPT) and DALL-E.
Overfitting : A phenomenon in artificial intelligence where a model adapts excessively to training data, capturing its specificities and noise, which harms its ability to predict or generalize on unseen data.
Peak Data: Peak Data refers to the moment when the amount of data available to train artificial intelligence models reaches its maximum, making the acquisition of new significant data increasingly difficult and potentially leading to errors.
Perceptron: An artificial neuron model capable of taking multiple inputs, weighting them, and then producing a binary output based on a threshold. It is one of the first supervised learning algorithms and the historical basis of modern neural networks.
Prompt: A prompt is an instruction or question given to an artificial intelligence system to generate a specific response or content. It serves as a starting point to guide the production of information or ideas.
Prompt Chaining: Prompt chaining is a technique consisting of linking several prompts to guide artificial intelligence in performing complex tasks or refining its responses. By using this method, more precise and consistent results can be obtained by structuring the process progressively.
Prompt Engineering: This is the art of designing and refining precise and structured instructions to interact with generative AI models, such as on Yiaho, in order to obtain relevant, coherent, or creative responses adapted to a specific objective.
Prompt Injection: A malicious technique aimed at manipulating an AI by inserting deceptive instructions into its inputs, to bypass its rules or obtain unauthorized responses.
RAG (Retrieval-Augmented Generation): An AI method that combines information retrieval and text generation. It allows a model to fetch external data (documents, knowledge base) to produce more accurate responses and avoid errors.
Reinforcement learning: Reinforcement learning is an artificial intelligence method where an agent learns to make optimal decisions by interacting with an environment, receiving rewards or penalties based on its actions.
Neural Network: A neural network is a computer model that mimics the functioning of the brain, consisting of interconnected neurons. It processes data by learning to recognize patterns through layers of neurons that transmit information between them.
Convolutional Neural Network: A CNN (Convolutional Neural Network) is a type of artificial neural network specialized in processing grid-structured data, such as images. It uses convolution layers to automatically extract features (edges, shapes, textures) and is particularly effective for tasks like image recognition or object detection.
Recurrent Neural Network (RNN): A type of artificial neural network designed to process sequences of data, with an internal memory that retains the context of previous information, ideal for language, music, or time series.
Backpropagation: An AI method that helps a neural network learn by correcting its errors. It adjusts internal connections to improve predictions by starting from the final error.
Sam Altman: Born on April 22, 1985, in Chicago, he has been the CEO of OpenAI since 2019 and the co-founder of this organization, which launched ChatGPT in 2022, a major breakthrough in AI. He also led Y Combinator from 2014 to 2019 and was briefly removed from OpenAI in 2023 before being reinstated.
AI Tests: AI tests are evaluations designed to assess the capabilities of artificial intelligences, such as their ability to imitate humans, solve problems, or create autonomously. Among the best known are the Turing Test, the Coffee Test, and the GAIA Benchmark, each revealing specific facets of machine intelligence.
Token: In artificial intelligence, a “token” is a unit of text that can be a word, a sub-part of a word, or even a character, depending on the tokenization method used. Tokens are used to transform raw text into a form that algorithms can understand and process.
Training Data: Training data in AI are datasets (inputs and expected outputs) used to teach an artificial intelligence model to recognize patterns, make predictions, or accomplish specific tasks.
Transfer Learning: In AI, Transfer Learning involves using a model pre-trained on a general task to accelerate and improve learning on a specific task, by transferring acquired knowledge.
Transformers: An artificial intelligence technology introduced in 2017, Transformers allow machines to understand and translate complex texts, generate images from simple descriptions, or create natural conversations, relying on an attention mechanism that analyzes the relationships between words or visual elements.
Tiny Recursive Models (TRM): Ultra-compact AI models (1 to 10M parameters) designed for structured reasoning with deep recursion and backtrack. They outperform LLMs in logic and math, while consuming 200 times less energy.
Underfitting : In artificial intelligence, underfitting occurs when a machine learning model is too simple to capture the complexity of the training data, leading to poor performance even on this initial data.
Vibe coding: An intuitive programming method where the user describes their ideas in natural language, and an AI generates the corresponding code. Popularized in 2025 by Andrej Karpathy, it prioritizes creativity and collaboration with online tools.
World model: A world model in artificial intelligence is a dynamic, learned internal representation that an AI system builds to model the physical, causal, and temporal laws of its environment. It allows for simulating future states, planning actions, and reasoning autonomously without constant interaction with the real world.
xAI: An artificial intelligence company founded by Elon Musk on March 9, 2023, focused on developing AI for scientific research.
Yann LeCun: French computer scientist, pioneer of artificial intelligence, known for developing the first convolutional neural networks (CNNs). In 1989, he created LeNet, a model for recognizing handwritten digits. Recipient of the 2018 Turing Award with Geoffrey Hinton and Yoshua Bengio, he is currently the Chief AI Scientist at Meta and a professor at New York University.
Zero-Shot Learning: Zero-shot learning is an AI technique that allows a model to recognize or perform a task on unseen data. This technique relies on semantic descriptions and transferable knowledge, without specific training.